Holodeck is my personal research project. It explores whether an application can be described so that an AI model executes its rules and generates the right interface for people and agents at the moment of use.
| Context | Personal project by Stefan Hamann |
|---|---|
| Status | Research project |
| Repository | github.com/sthamann/holodeck |
| Checked on | 18 November 2026 |
Deck, Holo and Host
It starts with a definition: objects, rules, rights and effects. Holodeck compiles them into a Deck. The Deck holds the executable meaning of the application, the evidence for it and the bindings for people and agents.
A Holo is an interface streamed live from the Deck. The Host supplies what a model cannot guarantee on its own: storage, transactions and time.
What the repository shows so far
These results come from the project documentation, where they can be reproduced.
- Seven formal planning tasks across Gripper, Blocks and Miconic compile through one pipeline.
- Weight modules built from four balls execute all 340 grounded actions in the 42 ball task.
- When a ball becomes protected mid-run, a new 116 step plan is produced. The constraint holds across commit, migration and restart.
- On a synthetic suite of 24 utterances, a frozen model made 248 of 288 decisions.
How it relates to commerce
Holoshop is the first experiment. It lets you replay a learned service credit rule with synthetic data from a second merchant. In the book, chapter 21 places Holodeck in the question of what business software looks like in a world of agents.